PantheonSim

Run CUDA and HIP programs, unmodified, on a CPU. No GPU needed. It pretends to be an NVIDIA or AMD GPU (a T4, an H100, an MI300X), so programs built with the real nvcc or hipcc run on it unchanged. It checks what your code does, not how fast it runs: there is no timing model, so a run on a physical GPU before a release still counts.

Try it in 60 seconds

With Docker (no GPU, no driver, no --gpus flag):

docker run --rm ghcr.io/pantheongpu/pantheonsim:latest vgpu list-gpus
docker run --rm -e VGPU_GPU=nvidia/h100 -e VGPU_DEVICE_COUNT=2 ghcr.io/pantheongpu/pantheonsim:latest nvidia-smi

The second command prints nvidia-smi output for two simulated H100s. The image has no compiler (nvcc and hipcc are not in it); see docs/docker.md to build a program inside a container, or use the source build below.

What it is: a CPU-side simulator of NVIDIA and AMD GPUs, so programs compiled with the real nvcc or hipcc run on a machine without a GPU. What it isn't: a performance model (timings mean nothing), and not every CUDA or HIP feature is covered; Status and limits lists what is.

Run it from source

make     # installs cmake and a C++ compiler if missing (asks first), then builds
./run    # starts a simulated GPU machine and takes you inside; type exit to leave

./run reads pantheonsim.yaml and writes one on the first run. To start from a ready-made file instead, use the sample: ./run pantheonsim.example.yaml runs it as it is (one simulated H100), and cp pantheonsim.example.yaml pantheonsim.yaml makes it your default to edit. It lists every setting, and "Choose your machine" below explains them.

You need Linux (WSL works), a C++20 compiler and CMake, and no GPU. Run both commands from the repository root: make: *** No targets specified or ./run: No such file or directory means you are in another directory or on an old branch, so git checkout main && git pull first. make doctor checks the tools and says what to install, make deps installs them, and make build builds without installing anything. The build takes a few minutes and uses half your CPUs (make JOBS=8 to choose). The simulator also runs on half your CPUs; set threads: N in the settings file, or make run THREADS=N, to change that. ./run before make builds for you.

Other targets:

make demo                           # run a built-in vectorAdd kernel on a simulated H100; ends with PASS
make run APP=./my_app ARGS="..."    # run one program on the simulated GPU, without the machine shell
make test                           # build and run the test suite
make help                           # every target and setting

Inside, nvidia-smi shows the GPUs, and you can compile, download programs with curl or git clone, and run them; for Python packages use a virtual environment (python3 -m venv ~/v && ~/v/bin/pip install ...); system packages (apt) are installed on your own machine, outside.

Choose your machine

Which GPUs: make list-gpus (or build/vgpu list-gpus) prints every id you can use, 48 NVIDIA and AMD cards from a T4 to a B300 and from an RX 6700 XT to an MI455X. Use the id as printed: gpu: nvidia/t4 in the settings file, or make run GPU=nvidia/t4. build/vgpu info --gpu nvidia/t4 shows one card. Some profiles were measured on a physical card and some come from public documents only; Status and limits says which.

Put the settings in a file, for example my.yaml, then ./run my.yaml. pantheonsim.example.yaml lists every setting and runs as it is (./run pantheonsim.example.yaml); copy it to pantheonsim.yaml to make it the default, which is also what ./run writes for you on a first run:

gpu: nvidia/h100     # which GPU (build/vgpu list-gpus shows them all)
count: 8             # how many
strict: true         # refuse undefined behaviour, such as dividing by zero

Every setting the file accepts (the full text for each is in Settings, and build/vgpu config keys prints it from the program):

Key	Values	Default	What it does
gpu	an id from list-gpus, such as nvidia/h100	nvidia/h100	The GPU model to present
count	1 to 16	1	How many identical devices
vram_mb	MiB, at least 1	the profile's own	Memory per device
cuda_version	major.minor, such as 13.0	this build's	The CUDA version nvidia-smi and NVML report
driver_version	such as 580.65.06	this build's	The NVIDIA driver version reported
rocm_version	such as 6.4.0	6.2.0 in the machine shell	The ROCm version amd-smi version reports
os	ubuntu, rocky or debian, with a version (ubuntu:24.04)	ubuntu:22.04	The operating system the machine presents
ram_mb	MiB, at least 1	unlimited	Host RAM device memory may use before it spills to memory_dir
memory_dir	a directory	never spill	Where device memory past ram_mb goes (runs at disk speed)
host_reserve_mb	MiB, 0 or more	6% of the limit, 1 to 4 GiB	Memory the guard leaves to the rest of your machine
host_guard	true or false	true	false turns the host-memory guard off
threads	1 to 65536	one per core (half under make run and ./run)	Host threads that run blocks
max_steps

In a browser: pantheonsim.com/play. In CI: pantheongpu/setup-pantheonsim@v0 (docs/github-action.md), or scripts/ci-setup.sh on GitLab CI, CircleCI, Forgejo, Jenkins and Buildkite (docs/ci-other-platforms.md). In a container: docker run --rm ghcr.io/pantheongpu/pantheonsim:latest vgpu list-gpus (docs/docker.md).

Docs

Getting started (step by step, your own programs, first-run fixes) | Settings | The machine | Compilers | Commands | Status and limits | Architecture | Contributing | Code of conduct | Governance | Security

Questions go in Discussions, bugs in issues. Apache-2.0 (see LICENSE); CUDA is a trademark of NVIDIA, and this project is not affiliated with NVIDIA or AMD.